Structural Prediction
in Incremental Dependency Parsing
Niels Beuck and Wolfgang Menzel
University of Hamburg
{beuck,menzel}@informatik.uni-hamburg.de
Abstract. For dependency structures of incomplete sentences to be
fully connected, nodes in addition to those corresponding to the words in
the sentence prefix are necessary. We analyze a German dependency corpus to estimate the extent of such predictive structures. We also present
an incremental parser based on the WCDG framework and describe how
the results from the corpus study can be used to adapt an existing
weighted constraint dependency grammar for complete sentences to the
case of incremental parsing.
Keywords: Parsing, Incrementality, Weighted Constraints, Dependency
Grammar
1
Motivation
A dependency structure of a sentence is a fully connected, single headed, directed
acyclic graph, where every node corresponds to a word in the given sentence.
For incomplete sentences, as encountered during incremental parsing, achieving
connectedness is often not possible when the dependency structure is limited to
the nodes corresponding to words in the sentence prefix. This problem occurs
for every word that is attached to the right, as there will be at least one sentence
prefix where the word but not its head is part of the prefix.
There are two possible approaches to deal with this problem: Either to drop
the connectedness requirement for the dependency structures of incomplete sentences (from now on called partial dependency analyses, PDA) or allow additional nodes in the structure as needed to establish connectedness. In the first
variant, words are left unattached until a suitable head appears. While this approach to incremental dependency parsing has been successfully used in parsers
like MaltParser [1], we argue that it is unsatisfactory when partial dependency
structures are of interest, for several reasons:
From a psycholinguistic point of view: Humans not only have interpretations of partial sentences, but also have strong expectations about the integration of words even if their syntactic head is still missing [2].
From an application point of view: Information already contained in the
perceived part of the sentence is missing in the dependency structure if
words are left unconnected. In the incomplete sentence “Pick up the red”,
the head of ‘red’ is still missing, but we can already argue that the object
of ‘pick’ is something red. This is not limited to words close to the end of
the prefix, like in a noun phrase under construction. It also applies to long
distance dependencies like in German subclauses, where the verb comes last
and all its arguments would need to stay unconnected. [3] proposed explicit
dependencies between the arguments, but these are unnecessary if we provide an extra node for the expected verb, allowing early attachment of the
arguments.
From a parsing perspective: Parsers choose between alternative structures
based on some measure of grammaticality or probability. The WCDG parser
used in this work uses a set of weighted constraints to measure the grammaticality of dependency structures. To be able to reuse measures optimized
for judging analyses of complete sentences, a structure similar to the final
structure is required to provide a rating similar to the rating for the final
structure. Otherwise, unintuitive structures with no possible counterpart in
a complete sentence analysis might be rated unreasonably high and lead the
parser astray.
On the one hand, additional nodes in partial dependency analyses are desirable
because they allow more informative structures. This leaves open the question
of how to construct such structures in a parser. On the other hand, a constraint
based parser would penalize partial dependency structures even if the constraint
violation could and probably will be remedied by further input, making it hard
to fairly rate and select a suitable dependency structure for a given sentence
prefix.
We propose that the second problem holds the answer to the first one. The
very constraint violations that unnecessarily penalize partial dependency analyses lacking structural prediction can be used as indicators for how to extend
these structures with the needed additional nodes.
The remainder of this paper consists of two parts: First, we will present a constraint based incremental parsing algorithm able to make structural predictions
about the upcoming input. It uses an existing weighted constraint dependency
grammar and an existing transformation based parsing algorithm as well as a set
of potential placeholders, so called virtual nodes (VNs), for structural prediction.
Secondly, we will analyze a given corpus of dependency annotations to assess the additional structure needed to keep PDAs connected and to fulfill all
valency requirements. From this result we can infer the set of VNs we have to
provide to our algorithm to achieve a certain coverage. In addition, we analyze
to what extend the given set of weighted constraints can be used to judge such
structurally predictive partial analyses and where it has to be adapted, e.g. by
adding suitable exceptions.
2
Related Work
While to our knowledge this is the first strictly incremental dependency parser
(strict in the sense of providing a fully connected structure for every input in-
crement), strict incrementality has been explored already for the Tree Adjoining
Grammar (TAG) formalism. [4] investigated the amount of structure needed for
strictly incremental parsing by building gold standard annotations incrementally
with a simulated parser. They introduced the term connection path to denote
the set of nodes of the (connected) tree for increment n that are not part of the
tree for increment n − 1 and not part of the subtree provided by the new word.
While they worked with phrase structures for English and only regarded predictions needed to achieve connectedness, our work is on dependency structures for
German and regards valency requirements in addition to connectedness.
PLTAG (PsychoLinguistically motivated Tree Adjoining Grammar)[5] is a
TAG variant, that allows strictly incremental parsing. There, connectedness is
facilitated by so called prediction trees, non-lexicalized elementary trees whose
nodes later need to be instantiated by nodes from lexicalized elementary trees.
Prediction trees are not anchored in the input directly, but are integrated when
necessary to connect the elementary tree of the new material to the structure
built so far. As the number of possible connection paths grows polynomially with
the number of prediction trees contained, the parser presented in [5] restricts
connection paths to one prediction tree.
The PLTAG parser deals with temporary ambiguity by keeping a beam of
possible derivation trees, ranked by their probability learned from a training
corpus. Thus, the best structure needs not be a monotonic continuation of the
previously best ranked structure. The output of the incremental WCDG parser
is non-monotonic, too, but instead of working with a beam of monotonically
extended structures, it works by successively transforming a single structure.
In contrast to the PLTAG parser, incremental WCDG achieves state of the art
accuracy (see [5], p.230 and [6]).
3
The WCDG Framework
Weighted Constraint Dependency Grammar (WCDG) [7] is a framework for dependency parsing by means of constraint optimization. The grammar consists
of a set of constraints regarding single or pairs of dependency edges. So called
context sensitive predicates allow constraints to access properties of the complete tree, e.g. to express the existence requirement for the subject of a verb.
The constraints are weighted with a penalty score between 0 and 1, where 0
denotes a hard constraint. By applying all constraints of the grammar to a dependency structure, a grammaticality score for the structure can be calculated
by multiplying the penalties of all constraint violations. Thus, parsing in WCDG
is the search for the highest scored dependency structure. Every word can, in
principle, be attached to any of the other words in the sentence, including the
generic root node via any dependency type. The only restriction is that a word
can only be attached to one head.1 All further restrictions, like projectivity, are
1
WCDG supports multiple levels, i.e. building multiple dependency structures at
once, including constraints mediating between them. Yet, for a given level, a word
can have only one head.
expressed via constraints. The number of possible edges grows quadratic with
the number words in the sentence, resulting in an exponential number of possible dependency structures. Thus, for non-trivial sentences, a complete search is
usually not feasible.
[7] presented a repair based heuristic algorithm called frobbing. The algorithm starts with an arbitrary initial structure and identifies candidates for
replacement by looking at the constraint violation with the highest penalty.
Replacement variants for the edges violating the selected constraint are then
evaluated for whether they would prevent this conflict and one of them is applied. Several of these repair steps might have to be chained to escape from a
local score maximum.
This algorithm is inherently non-incremental, as it starts with a structure
for the whole sentence. An incremental processing mode can be introduced by
applying the procedure on prefixes of the input and using the resulting dependency structure as the initial structure for analyzing the extended prefix. In this
incremental mode, edges selected in a previous step can be replaced later on, if
necessary. This can happen when, given the additional input, they violate additional constraints or when an old constraint violation can be resolved given the
new input. The algorithm guarantees no monotonicity at all, but in practice an
attachment stability of around 70% has been observed [6].
4
Structural Prediction with Virtual Nodes
When parsing a sentence prefix instead of a full sentence with WCDG, several
constraints might prevent the parser from choosing a prefix of the structure
which would be selected for the full sentence, i.e. a structure only containing dependency edges also present in the complete structure. This is foremost due to
‘fragmentation’ constraints, penalizing unattached words. They force the parser
to select an unsuitable attachment point that might be penalized by other constraints, but not as hard as the fragmented reading. Also, valency constraints
are commonly violated in such prefix structures.
There are two possible approaches to deal with these kinds of constraints
when applied to partial dependency analyses: Either we prevent them from being applied to PDAs or we add as many additional nodes and edges to the
dependency structure as needed to prevent the constraint violation. The first
approach has the severe drawback of over-generation. Disabling constraints (or
adding exceptions to them) will remove their ability to penalize ungrammatical
input.
In this paper we follow the second approach by providing a set of so called
virtual nodes that can be integrated into the dependency structure if a constraint violation can be avoided by doing so, but there is no penalty if they stay
unconnected. If being unconnected, they are not considered to be part of the
structure.
To prevent the parser from choosing a structure that includes virtual nodes
even if another structure of equal or slightly worse score would be available, we
S
S
He
DIROB
J
BJ
BJ
SU
SU
a
buys
a
buys
He
VS
DE
T
[noun]
S
Fig. 1. Example for bottom-up prediction: The determiner is unattachable to any word
directly, but attachable if predicting a noun
OB
BJ
Er
sieht
[virtNoun]
PP
ET
JA
SU
PN
D
; Das
Haus
mit
[virtNoun]
Fig. 2. Examples for top-down predictions of an object and the kernel noun of a preposition (“he sees” and “the house with”)
add a slight penalty to every analysis where a virtual node is used, i.e. connected. This way, the structurally predictive reading is chosen only if necessary,
but not otherwise. No further modification to the algorithm is needed. During
constraint optimization the virtual nodes are treated like other input nodes. The
special handling of virtual nodes (i.e., the slight penalty for connected and no
penalty for unconnected virtual nodes, in contrast to the harsh penalty for other
unconnected nodes) is defined solely in the constraints of the WCD-Grammar.
We can distinguish two different scenarios leading to structurally predictive
readings. The scenario seen in Figure 1 is a bottom-up prediction, where the
necessity to connect a new word leads to the prediction of a suitable head.
As seen in Figure 2, dependents can be predicted, too, via top down prediction. Whenever a word has a valency that can not be satisfied by any of the other
available word, e.g. missing subject or a transitive verb missing an object, filling
the valency with a VN makes the corresponding constraint violation disappear.
If we would allow some words to stay unattached, it cannot be guaranteed
that there is any possible continuation of the sentence, where these unattached
words could be attached without changing the already established part of the
structure. For fully connected structurally predictive structures, in contrast, the
absence of severe constraint violations licenses not only the integration of the
otherwise unattachable words but also the acceptability of the whole structure.
It anticipates a potential continuation of the sentence prefix, which does not
introduce any additional constraint violations. The structural prediction serves
as a kind of “certification” for the grammaticality of all the attachments in the
PDA.
It has to be assured that for every sentence prefix the grammar licenses a
partial dependency analysis free of additional constraint violations. Besides providing a suitable set of virtual nodes, grammar modifications might be necessary
to achieve this goal, as will be discussed in Section 6.
To be able to predict structural relationships for yet unseen words, the incremental frobbing algorithm needs to be provided with a finite set of virtual nodes.
The larger this set, the higher the coverage will be. Additional VNs, however,
introduce a quadratic increase of potential dependency edges into the search
space and, as a consequence, an exponential increase in the number of possible
dependency structures.
For the sake of computational feasibility, the number of VNs should be kept
as small as possible. This can be achieved by using general VNs (which stand
for whole classes of words such as nouns or verbs). Aside from the combinatoric
problems of providing too many VNs, it also avoids the danger of many variants receiving the same grammaticality score and thus being indistinguishable
by the parser. Even the order among VNs will remain unspecified. While the
existing grammar is prepared to deal with underspecified values for many lexical
features, attributes like the word form or the linear precedence relations among
words are always considered being given explicitly. To allow constraints to ignore
the arbitrary order/adjacency relations between VNs as well as their particular
lexical instantiations, respective exceptions need to be added.
Thus, the questions we have to answer are:
– Which set of virtual nodes is needed to achieve a certain coverage
– What changes to a constraint dependency grammar for full sentences are
necessary?
In the second half of the paper we will present an empirical investigation to
answer these questions.
5
Data Preparation
Not only we have no annotations for sentence prefixes available, but also it would
be difficult to specify them uniquely, since they always depend on a hypothesis
for a possible sentence continuation. Therefore, we generate them from complete
sentence annotations. We want to find a monotonic sequence of prefixes resulting
in the final annotation. Those prefixes are therefore not necessarily the most
plausible for any given prefix.
For every word in a given sentence, we generate an annotation up to that
word by retaining certain nodes and dependency edges while dismissing others2 .
5.1
Prefix Baseline Annotation
The basic variant of the procedure that generates connected PDAs is quite
straightforward:
– all words inside the prefix and dependencies between them are kept
– all words that are heads of retained words are retained themselves, including
the dependency edge between the head and the dependent. This rule applies
recursively.
2
The software to reproduce this data preparation as well as the evaluation can be
found at http://www.CICLing.org/2013/data/274
S
J
SUB
DET TR
AT
DET TR
AT
Der
kleine
[virtNoun]
[virtVerb]
VS Der
kleine
[virtNN]
Fig. 3. Example for assuming a complete sentence vs a fragment given the prefix “der
kleine” (“the little”)
KONJ
SUBJ
X
AU
KONJ BJ A
SU BJ
O
X
AU
A
OBJ
Dass
er
ihn
[VVPP]
[VAINF]
[VMFIN]
⇒ Dass
er
ihn
[VVFIN]
Fig. 4. Example for folding: When generating a PDA for the first three words of the
clause “dass er ihn gesehen haben soll” (literally: “that he him saw has should” meaning:“that he is supposed to have seen him”) . All three verbs would be kept for connectedness, but are then folded into a single virtual verb
Words not belonging to the prefix but nonetheless retained need to be ’virtualized’ to produce generic representatives of their part-of-speech category, the
form of the virtual nodes used by the incremental parsing algorithm.
The procedure of virtualization removes word form and lemma, keeps partof-speech information, removes all other lexical features, and blurs the original
linear order among the nodes.
5.2
Prefix Annotation Variants
Three parameters for the prefix generation algorithm have been considered:
1. Assuming complete sentences or not, i.e. whether we retain all nodes needed
to connect up to the root or only retain those needed to connect all words
in the prefix to each other (see Figure 3).
2. Top-down prediction. The connectedness criterion only provides for bottomup prediction. Additional nodes could be retained via top-down prediction,
i.e. to fill obligatory valencies (see Figure 2).
3. Folding of connection paths. The connection paths extracted directly from
the full sentence annotations might not represent the shortest possible connection path. This is especially true for chains of equally labeled dependencies which correspond to nested phrase structures. Folding these structures
by only keeping one of the dependencies will lead to more plausible prefix annotations (see Figure 4). They correspond to the minimally recursive
structures [4] investigated in their second experiment where they anticipated
parser miss-analyses in their simulated parsing procedure.
These three parameters have different consequences for the parser. Unless there
are clear hints for a nested structure, the minimal tree, i.e. the one with the
least number of virtual nodes, should always be preferred. WCDG achieves this
due to the prediction penalty, selecting the PDA with a minimal amount of
VNs among otherwise equally scored alternatives. The parser will not be able
to utilize additional VNs set aside for additional nesting depth. Folding can be
expected to provide a better estimation of the amount of VNs needed.
While an incremental dependency parsing algorithm that provides bottom-up
but no top-down prediction is possible, the parsing algorithm used here has no
distinct mechanisms for the two prediction modes. If respective virtual nodes are
provided, they will be used to satisfy valency constraints, unless those constraints
are disabled for incomplete input. Therefore, by adding top-down prediction to
the prefix generation algorithm, we get a better estimation for the VNs needed
by the parser. Note that folding reduces the number of VNs, while top-down
prediction increases it.
In contrast to folding, where there is no choice for parser behavior, the parser
can easily be changed between a full sentence and a fragment mode. Which mode
is activated depends solely on whether there is a constraint for penalizing words
other than finite verbs to be the root of the tree. Such a constraint would incite
the prediction of a virtual finite verb (and possibly its subject) even if none
is needed to establish connectedness. Thus, the choice between these modes
affects the number of required virtual nodes. In this paper we only explored the
fragment variant, as did [4] and [5]. The predictions in this mode are expected
to be smaller and less speculative.
6
Experiments
To determine the set of virtual nodes needed, we generate a partial but connected
structure for every prefix of every sentence in an annotated corpus. A subset of
the dependency edges of the complete tree is selected by using the procedure
presented above. Given these structures, we can count how many additional
nodes are needed to keep them connected. Also, by evaluating these PDAs with
the WCD-grammar, we can determine whether they violate any constraints in
addition to those violated in the complete sentence, as such additional constraint
violations would prevent the prefix structure to be selected by the parser. Based
on these results we can then modify the grammar accordingly.
Using the 500 sentences from the Negra corpus [8] transformed to the dependency annotation scheme described in [9] and a broad coverage WCD-grammar
for German [7] we perform the following procedure:
1. generate a set of prefix annotations from the gold standard annotation.
2. Calculate constraint violations that only appear on prefix analyses, but not
in the corresponding complete sentence
3. check:
(a) whether changes to the constraints themselves are needed
(b) whether additional VNs corresponding to words in the complete sentence
could avoid any of the additional constraint violations
(c) whether VNs can be removed from the connection path systematically
without triggering additional constraint violations
4. repeat from step 1 with different variants of the PDA generation algorithm
6.1
Prediction with Baseline Annotations
We start with the most simple variant of the prefix generation algorithm, i.e.
without folding or top-down prediction. The constraint violations encountered
in the prefix annotations but not in the respective complete sentence annotation
can be roughly categorized as follows:
Prediction penalties As described in Section 4, there is a small penalty for
integrating VNs into a structure. The respective constraint violations are
expected in the difference lists, and can safely be ignored here.
Order constraints, especially projectivity These constraint violations occur due to VNs being unordered. These constraints should simply not be
applied to VNs. Hence, respective exceptions need to be added to them.
Unsatisfied valencies As we did not account for valencies when deciding on
what to keep in the prefix structures of this iteration, all kinds of valency
constraints are violated, e.g. missing subject, missing object, missing kernel
noun for prepositions, missing determiners for nouns, missing conjunctions
and missing conjuncts. These are prime examples for top-down prediction,
i.e. these conflicts will disappear when attaching a VN with a respective PoS
and dependency type. For VNs missing determiners or infinitive markers
‘zu’, PoS variants without these valencies (NE and VVIZU respectively) will
be selected instead of top-down-prediction.
Punctuation Several constraints expect a comma or quotation marks between
two words. If at least one of the words is virtual, the missing punctuation
might still show up later on. Thus we will add respective exceptions to these
constraints.
Word form There are constraints accessing the word forms or lemmas. Many
of these occurrences are exceptions for certain words. We will ignore them for
now, as we neither want too detailed predictions (exploding the search space
of the parser) nor do we want to allow idiosyncratic constructions before the
licensing word appears in the input. For constraints deemed general enough,
exceptions which prevent them to be applied to VNs are added.
Table 1 shows the percentage of prefix structures containing a certain number
of VNs, in total and coarsely grouped by part of speech. We also added a line for
the number of VNs in the connection path, i.e. only the VNs needed in addition
to those already present in the previous prefix, to connect the most recent word
to the rest of the structure. This line is roughly comparable to the numbers
presented in [4]. Their headless projections in the connection path correspond
to our VNs. In both cases connection paths up to length 1 cover 98% of the
observed prefixes and 2 cover over 99%. No occurrences of path longer than 4
are observed. Instead of 82.3% of prefixes not requiring any headless projections
there, 80.8% of the words can be connected without VNs here.
To estimate whether the minor difference results from differences in language
(English VS German) or grammar formalism (phrase structure VS dependency
structure), we ran our evaluation on 500 sentences from the WSJ corpus, i.e.
Table 1. Results of experiment 1: The number of VNs needed for prefix structures,
only regarding connectedness but no valencies; total VNs: number of VNs per PDA;
CP length: only those VNs needed to connect the most recent word; note that the
numbers in the columns don’t add up: the 4 VNs of a PDA might consist of 3 virtual
verbs and 1 virtual nominal
0
1
total VNs
verbs
nominals
adjectives
other
CP length
56.5%
68.8%
81.2%
97.5%
99.1%
80.8%
WSJ
79.4% 17.8%
2
3
4 5+
30.1% 10.8% 2.4% 0.2% 0%
26.0% 4.6% 0.7% < 0.1% 0%
18.7% 0.1% 0%
0% 0%
2.5% < 0.1% 0%
0% 0%
0.8% < 0.1% 0%
0% 0%
17.0% 2.0% 0.1%
0% 0%
2.7% 0.2%
0% 0%
the same corpus used in [4], converted to dependency structures. The results (cp
length only) are shown in the last line of Table 1 and are closer to our results,
suggesting that the differences result from the grammar formalism.
6.2
Valency Respecting Prediction
In the previous experiment, a large class of constraint violations were related to
unsatisfied valencies. We therefore modified the prefix generation algorithm to
keep words in the structure for the following cases:
– Subjects of all finite verbs, virtual or not
– Kernel nouns of prepositions
– Dependents of a conjunction
– Objects of non-virtual words
– The second part of a truncation like ‘an- und verkaufen’ if the first is observed
– The last part of a conjunction chain, if the first two are observed, e.g. in the
sequence ‘A, B, C und D’, ‘und D’ would be kept and C skipped, if ‘A,B’
was observed
– Certain adverbs that fill the role of conjunction in subclauses are kept, if the
verbal head of the subclause is already kept for other reasons.
Since the valencies of virtual words are not available to the parser, they all have
to be considered optional. Therefore, no objects for virtual verbs are kept.
6.3
Prediction with Folded Structures
The results for PDAs with folding are shown in Table 3. Compared to Table 1,
the most noticeable difference is that the percentage of prefixes with two virtual
verbs went down by around 3.5%, while the number for 1 virtual verb increased.
As expected, there is no change in the zero column number, as folding will not
remove the last virtual verb remaining.
Table 2. Results of experiment 2: The number of VNs needed for prefix structures,
regarding connectedness and valencies
total VNs
verbs
nominals
adjectives
conjunction
adverb
preposition
other
0
1
37.4%
62.6%
58.5%
96.1%
97.9%
99.4%
98.3%
99.2%
34.6%
30.8%
37.5%
3.8%
2.1%
0.4%
1.7%
0.9%
2
3
4
5+
20.7% 5.8% 1.3% 0.2%
5.8% 0.8% < 0.1% 0%
3.9% 0.1%
0% 0%
0.1% 0%
0% 0%
0% 0%
0% 0%
0.2% 0%
0% 0%
0% 0%
0% 0%
0% 0%
0% 0%
Table 3. Results of experiment 3: The number of VNs needed for prefix structures,
regarding connectedness while folding recursive verb chains
0
total VNs
verbs
nominals
adjectives
other
CP length
6.4
56.5%
68.8%
81.2%
97.5%
99.1%
81.15%
1
2
3
4 5+
33.3% 9.3% 0.9% < 0.1% 0%
30.2% 1.0% 0.1%
0% 0%
18.8% 0.1%
0%
0% 0%
2.5% < 0.1%
0%
0% 0%
0.8% < 0.1%
0%
0% 0%
17.2% 1.6% < 0.1%
0% 0%
Prediction with Valencies and Folding
The results for the combined algorithm, folding plus valencies, can be seen in
Table 4. As expected, the numbers are generally above those of Table 3 and
below those of Table 2. Three VNs are sufficient to achieve a coverage of 99%.
One line in the table that might surprise is that prepositions as VNs were
observed (1.7%), since a preposition is the leftmost word in a PP. There are three
different ways prepositions are predicted: to connect an adverb (arguably a very
ambiguous and hard to predict case), to fill a verb’s valency for a prepositional
object, and to complete a coordination of prepositional phrases, when the first
PP and a conjunction like ‘und’ was already observed.
The question remains, what part-of-speech category combinations among the
VNs achieve which coverage. Some possible combinations are shown in Table 5.
A set consisting of two nominals and one verb, as used in [6], covers nearly
90%. Adding a VN for adjectives, conjunctions, prepositions and adverbs improve the coverage towards 99%. Whether the parser can actually benefit from
this improved coverage, especially of the latter two categories, will have to be
evaluated, but is beyond the scope of this paper.
Table 6 shows how VNs are distributed over the different prediction constellations. If a VN serves to connect two different words, e.g. it is the head
of a determiner the object of a verb, the leftmost word is selected. The most
common reason for prediction is predicting the kernel noun of a PP (18%), followed by predicting the subject of a finite verb (11%), predicting a clause initial
conjunction (10%) and connecting a determiner (6%).
Table 4. Results of experiment 4, the complete algorithm: The number of VNs needed
for PDAs, regarding connectedness and valencies while folding nested structures
total VNs
verbs
nominals
adjectives
conjunction
adverb
preposition
other
0
1
37.4%
62.6%
58.5%
96.1%
97.9%
99.4%
98.3%
99.1%
36.7%
35.0%
37.5%
3.8%
2.1%
0.4%
1.7%
0.9%
2
3
4
5+
20.6% 4.3% 0.9% 0.1%
2.3% 0.1% 0% 0%
3.9% 0.1% 0% 0%
0.1% 0% 0% 0%
0% 0% 0% 0%
0.2% 0% 0% 0%
0% 0% 0% 0%
0% 0% 0% 0%
Table 5. Percentage of sentence prefixes that can be represented with a given set of
VNs. N: nominal, V: verb, A: adjective, C: conjunction, P: preposition, Av: Adverb
VN-Set
no VNs
2∗N +V
2∗N +V +A
2∗N +2∗V +A
2∗N +2∗V +A+C
2∗N +2∗V +A+C +P
2 ∗ N + 2 ∗ V + A + C + P + Av
6.5
coverage (%)
37.4
89.4
93.0
94.7
96.7
98.3
98.7
Discussion
This study only covers part of the problem of grammar development: We made
sure for every prefix of every sentence there is a dependency structure that is
scored at least as good as the dependency structure for the final sentence. So far,
however, we did not consider the complementary case, namely to make sure that
for the best scored partial dependency structure of every sentence prefix there
actually is a continuation that would indeed be scored as good. This cannot be
examined by a corpus study but only by inspecting actual parser behavior.
7
Conclusions
In this paper we presented a parsing system based on the WCDG framework
capable of incrementally parsing into connected dependency structures. The system uses a grammar consisting of weighted constraints as well as a set of virtual
nodes. While there is an existing state of the art broad coverage WCD-grammar
for German, that grammar is optimized for complete sentences, not for sentence
prefixes. We conducted a corpus study to identify what changes to the existing
grammar are needed, as well as to identify sets of virtual nodes suitable to cover
the observed sentence prefixes. Our experiments complement those conducted
by [4] by exploring another grammar formalism and valency driven top-down
prediction in addition to connectedness-driven bottom-up prediction.
Table 6. An overview over common reasons for why VNs where included in a PDA;
bottom-up prediction to the left, top-down to the right; determined for the PDAs of
experiment 4;
reason
reason
connecting a
- determiner
- subject
- clause initial conjunction
- adverb
- accusative obj.
- preposition
% of VNs
6.0
5.9
5.6
4.5
3.8
3.4
filling the valency:
- kernel noun of a prep.
- subject
- the conjunct under
a conjunction
- accusative obj
- conjunction to finish
an incomplete coord.
- auxiliary verb
- predicate
- subject clause
- object clause
% of VNs
18.0
11.4
10.3
5.3
3.5
2.7
3.4
2.7
1.1
References
1. Nivre, J., Hall, J., Nilsson, J., Chanev, A., Eryigit, G., Kübler, S., Marinov, S.,
Marsi, E.: Maltparser: A language-independent system for data-driven dependency
parsing. Natural Language Engineering 13 (2007) 95–135
2. Sturt, P., Lombardo, V.: Processing coordinated structures: Incrementality and
connectedness. Cognitive Science 29 (2005) 291–305
3. Bornkessel, I.: The Argument Dependency Model: A neurocognitive approach to
incremental interpretation. Max Planck Institute of Cognitive Neuroscience (2002)
4. Lombardo, V., Sturt, P.: Incrementality and Lexicalism: a Treebank Study. In:
Lexical Representations in Sentence Processing. John Benjamins (2002) 137–154
5. Demberg, V.: A Broad-Coverage Model of Prediction in Human Sentence Processing. PhD thesis, The University of Edinburgh (2010)
6. Beuck, N., Köhn, A., Menzel, W.: Incremental parsing and the evaluation of partial dependency analyses. In: DepLing 2011, Proceedings of the 1st International
Conference on Dependency Linguistics. (2011)
7. Foth, K.A.: Hybrid Methods of Natural Language Analysis. PhD thesis, Uni Hamburg (2006)
8. Brants, T., Hendriks, R., Kramp, S., Krenn, B., Preis, C., Skut, W., Uszkoreit, H.:
Das negra-annotationsschema. Negra project report, Universität des Saarlandes,
Computerlinguistik, Saarbrücken, Germany (1997)
9. Daum, M., Foth, K., Menzel, W.: Automatic transformation of phrase treebanks
to dependency trees. In: 4th Int. Conf. on Language Resources and Evaluation,
LREC-2004. (2004)